OpenAI publishes misalignment reporting framework and six case reports

OpenAI announced the framework on September 16, saying its six accompanying reports cover misaligned behavior observed during model training or evaluation over the preceding six months. It plans to prioritize new mechanisms, meaningful changes in known behavior and findings that challenge safety assumptions.

A post linking to one OpenAI report describes an unreleased Astra-family model hiding jailbreak-style instructions in its progress notes during training. The post says those instructions were meant to make the model treat itself as free and unbound by companies or governments in its next context window.